Robust Feature Similarity Learning Framework Enhances Deepfake Detection Under Adversarial Attacks

Researchers S. Suriya and I. N. Sountharia have introduced a Robust Feature Similarity Learning framework designed to enhance deepfake detection under adversarial conditions. The proposed system improves detection reliability and feature discrimination by combining convolutional neural feature extraction with algorithmic parameter optimization and dedicated defense mechanisms.

The study addresses the critical challenge deepfake technology poses to the authenticity of digital media. Manipulated videos frequently appear authentic and are capable of evading standard media detection tools when subjected to adversarial perturbations.

To process visual inputs, the framework utilizes a Convolutional Artificial Neural Network to extract discriminative spatial feature embeddings. In parallel, the Logistic Chaotic Grey Wolf Algorithm optimizes the feature-selection mask and loss-balancing coefficients throughout the training phase.

The architecture also incorporates an Adversarial Feature Contrast Mechanism to differentiate between real and fake samples across different classes. In addition, a Perturbation-Aware Similarity Loss maintains similarity consistency between uncorrupted media samples and those modified by adversarial perturbations.

During model inference, the framework applies a Feature Squeezing Defence Mechanism to minimize the influence of nuisance perturbations.

In experimental evaluations conducted on the Celeb-DF (v2) benchmark dataset under normal conditions, the proposed framework achieved a detection accuracy of 97.50%. When tested under adversarial attack conditions, baseline accuracy dropped to 50%, but the integrated defense mechanism restored performance to 94%, confirming the model’s resilience and adversarial stability.

The comparative analysis detailed in the research demonstrates that the primary advantage of the framework stems from the systematic integration of its core components: the Logistic Chaotic Grey Wolf Algorithm, the Adversarial Feature Contrast Mechanism, the Perturbation-Aware Similarity Loss, and the Feature Squeezing Defence Mechanism. According to the findings, combining these distinct algorithmic strategies produces measurable enhancements across three key performance dimensions: feature discrimination, adversarial stability, and detection reliability.

Subject of Research: Deepfake detection using robust feature similarity learning

Article Title: Robust feature similarity learning framework: enhancing DF detection using CANN and LCGWA

Article References: Suriya, S., & Sountharia, I. N. (2026). Robust feature similarity learning framework: enhancing DF detection using CANN and LCGWA. Neural Computing and Applications, 38(21), Article 868. https://doi.org/10.1007/s00521-026-12637-4

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12637-4

Keywords: deepfake detection, feature similarity learning, neural networks, adversarial robustness, Celeb-DF, spatial feature embeddings, optimization algorithms, digital media authenticity, Robust, feature, similarity, learning